Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 29, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash is a 9 billion parameter language model distilled from DeepSeek-V4, leveraging the Qwen3.5-9B architecture. It is specifically designed for enhanced structured reasoning, multi-step problem-solving, and reliable agentic action generation, while maintaining efficient flash inference. This model excels at transferring advanced logical capabilities into a smaller, faster framework, making it suitable for complex reasoning tasks and tool-augmented workflows.

Loading preview...

Model Overview: Qwen3.5-9B-DeepSeek-V4-Flash

This model, developed by Jackrong in collaboration with hardware engineer Kyle Hessling, is a 9 billion parameter language model distilled from the high-quality reasoning data of DeepSeek-V4. It leverages the efficient Qwen3.5-9B architecture to transfer advanced structured reasoning and multi-step problem-solving capabilities, typically found in larger models, into a more agile framework. The distillation process, using the Jackrong/DeepSeek-V4-Distill-8000x dataset, focuses on capturing genuine logical generalization rather than superficial chain-of-thought.

Key Capabilities & Design:

  • Structured Reasoning: Inherits DeepSeek-V4's deep logical capabilities, enabling complex problem-solving.
  • Flash Inference: Optimized for speed and token-efficiency within its 9B parameter size.
  • Tool-augmented Workflows: Designed for reliable agentic action generation and interaction with tools.
  • Distillation Insights: Activates latent knowledge within the Qwen3.5-9B model and teaches actual problem-solving procedures, not just output formats.

Performance & Evaluation:

Early controlled evaluations, conducted by Kyle Hessling, show improved performance in agentic reasoning, front-end design, and tool calling compared to the official Qwen3.5-9B base model. This is attributed to the high-quality long-CoT data from DeepSeek-V4, which enables cross-domain transfer of reasoning abilities.

Best Practices:

For optimal results, use temperature=0.7 to 1.0 (lower for coding, higher for creative reasoning) and top_p=0.95. Employing a structured prompt template or standard ChatML format is recommended for best reasoning outcomes.

Limitations:

Despite its enhancements, the model is still constrained by its 9B parameter size, potentially struggling with extremely obscure knowledge. It may also exhibit "over-reasoning" on very simple queries due to its fine-tuning bias, and some alignment-sensitive behaviors might regress due to the focus on reasoning gains.